IWAISAKI Department of Animal Science, College of Agriculture Kyoto University, Kyoto-shi 606, Japan Summary The objective of this study was to find, using numerical methods, a configura
Trang 1Constructing categories for polychotomous variates
H IWAISAKI Department of Animal Science, College of Agriculture Kyoto University, Kyoto-shi 606, Japan
Summary
The objective of this study was to find, using numerical methods, a configuration of response categories which maximizes the ratio of heritability in a discrete scale to that in the underlying
normal scale for polychotomous characters It is shown that to maximize this ratio it is necessary
to construct a symmetric discrete distribution with symmetrically spaced weights for the categories
of response The thresholds that maximize the ratio are located at the arithmetic mean of the
expected values, on a normal scale, of contiguous categories The value of the ratio for the
a Optimal discrete distribution found in this study is compared with those for some other discrete distributions
Key words : Categories, polychotomous variates, heritability.
Résumé Construction de catégories pour des variables polychotomiques
Le but de cette étude est d’établir, par des méthodes numériques, une configuration des
catégories de réponse qui maximise le rapport de l’héritabilité d’un caractère polychotomique
définie sur une échelle discrète à celle définie sur l’échelle normale sous-jacente On montre que,
pour maximiser ce rapport, il est nécessaire de construire une distribution discrète symétrique,
avec des scores répartis symétriquement dans les diverses catégories de réponse Les seuils qui maximisent le rapport considéré sont égaux aux moyennes arithmétiques des espérances mathéma-tiques, sur une échelle normale, des catégories adjacentes On compare la valeur du rapport des héritabilités correspondant à la distribution discrète optimale trouvée dans cette étude à celles
correspondant à quelques autres distributions discrètes
Mots clés : Catégories, variables polychotomiques, héritabilité
I Introduction
Threshold characters present a discrete phenotypic distribution and heritability on
such a scale depends on the population frequencies in the different categories of
response (D & L , 1950 ; V et al., 1976 ; G IANOLA , 1979) As far as
Trang 2phenotypes concerned, it is already known that the ratio of
heritability in a discrete scale to that in the underlying normal scale is maximum with
an incidence of 0.5, and that the maximum expected heritability in a discrete scale is about 65 p 100 of that in the underlying normal scale (e.g., FALCONER, 1981 ; JAMES &
M
, 1982) This finding, in other words, implies that the dichotomous distribution with a 50 p 100 incidence is a discrete distribution which minimizes the loss in
heritability to that of the underlying normal variate In this context, however, no such theoretical evidence is available for cases of polychotomous characters with 2 or more
thresholds
Thus, the objective of this study was to evaluate the maximum of the ratio of
heritability in the discrete scale to that in the underlying normal scale and to find the
configuration of a discrete distribution which maximizes the ratio
II Methodology
G (1979) has developed the following general formula relating heritability in the narrow sense in the underlying normal scale to that in the outward scale :
where k is the heritability in an outward scale ; h is the heritability in the underlying
normal scale ; m is the number of mutually exclusive and exhaustive categories of
m
response ; ’1TI is the probability of response in the i’&dquo; category with _I ’1T¡ = 1 ;
i=l
t is the i’&dquo; fixed threshold with 4! (t ) =
(2’1TtII2 exp (- t?/2) ; a is the weight for the i’&dquo; category.
Eq (1) can be used for this study in the form of F (t, a) = hV! The problem is
to find the maximum of F (t, a) with respect to t’ =
[t&dquo; t!, , t _,] and a’ = [a,, a!, , a , and maximizing F (t, a) with respect to the unknown parameters
requires solving :
From (2) the following system of equations is derived :
Trang 3On the other hand, without loss of generality (3) with !, a1T¡ = 0 leads to :
i=l
as has been stated already by G & N (1981) Hence, using I aTr, = 0
i=l
and (5) and rearranging (4) gives :
Now we take 3 successive standard normal variates t 11 t!, t (t < t! < t ,) and their symmetrical points about zero — t - t , -
t (— t -, > — t > — t )- So, from
one can find sequentially if G (t ) = 0, also G (— t k ) = 0 for k = 1, 2, , m — 1 and
consequently from (5) one can also find ak =
a for k = 1, 2, , m These facts show that a necessary condition to maximize hyk is to take thresholds to form a
symmetric distribution with symmetrically spaced weights around the mean The vector
t maximizing h3hfl would have elements consisting of the arithmetic means between ak and ak&dquo; ; k = 1, 2, , m — 1 The procedure developed by BRENT (1973) was used to
find an iterative solution to the system (6) The procedure is a modification of Newton’s method which utilizes orthogonal transformations Equal frequency values for response categories were taken to obtain initial t’s for iteration The calculation of 1£;
was made on the basis of the method of CLENSHAW & CuRns (1960).
Values of hyh§ are given in table 1 for the following cases : 1) « optimal distribu-tion » (as given in this paper) with either normal or equally spaced scores, 2) equal percentile distribution, with normal or equally spaced scores, and 3) equally spaced
thresholds with normal or equally spaced scores Computation was performed by
FACOM M-382 at the Data Processing Center of Kyoto University.
III Results and discussion
Table 1 presents the value of hyk in the situation of the « optimal distribution »
with normal scoring of categories for a given number of categories, in comparison with those in other natural situations For dichotomous responses, the maximum expected
ratio of hyh§ is well known to be only 0.637 This work reveals that the maximum of the ratio in the trichotomous case is about 0.80 with an increment of 0.17 from that in the dichotomous and the maximum value in the pentachotomous case is above
Trang 40.90 The maximum, always 1, the number of response categories
increases When the number of categories is as many as 10, the outward variable
following the optimal distribution found may be as heritable as the underlying variable
Differences in hyk between the optimal distribution and some distributions
investi-gated in this study were found to exist although the magnitudes were small on the whole In cases with less than 5 categories, the value of h!h! for the equal percentile
distribution is closer to that for the optimal distribution and higher than that for the distribution derived from equally spaced thresholds, and vice versa in the cases of more
than 5 The 2 different scoring systems for categories, normal scoring and equally spaced scoring, had very little effect on the ratio value in all the situations as has been found by G & N (1981) This result is due to the fact that the differences between scores of adjacent categories are approximately the same with the 2 scoring
procedures.
Trang 5probabilities of response for categories yield h!/
1!; with normal scoring of categories The distributions show a bell-shaped symmetrical
pattern as would be expected These are the discrete distributions which fit « best » a
N (0, 1) for a given number of categories, and in this situation of frequency distribution with normal scoring of categories, response to selection would be maximum
In conclusion, the results obtained in this work suggest how to construct a discrete distribution from data on a normal one with the restriction of minimizing the loss in
heritability for a given number of categories The answer is that the discrete distribu-tion should be symmetric (table 2) and that categories should be scored using normal
scores This finding may be of interest when a categorical presentation might be better
adapted than a continuous one for practical reasons and may also be useful in studies
measuring the loss of information occurring from rendering a normal distribution into a
discrete one.
Received July 16, 1984
Accepted June 26, 1985
Acknowledgements
The author wishes to thank the 2 referees and a scientific editor, and Prof Y Y AMADA for
their valuable suggestions on the manuscript.
References
BRENT R.P., 1973 Some efficient algorithms for solving systems of nonlinear equations SIAM J Numer Anal., 10, 327-344
C C.W., Cuis A.R., 1960 A method for numerical integration on an automatic computer Numer Math., 2, 197-205
D E.R., L I.M., 1950 Heritability of threshold characters Genetics, 35, 212-236.
FALCONER D.S., 1981 Introduction to Quantitative Genetics, 2nd ed., 275 pp., Longman, London
G D., 1979 Heritability of polychotomous characters Genetics, 93, 1051-1055
G D., N H.W., 1981 Scaling threshold characters Genetics, 99, 357-384
JAMES J.W., Mc GUIRK B.J., 1982 Regression of offspring on parent for an all-or-none trait Z. Tierzuchtg Zuchtgsbiol., 99, 308-314
VW.E., WHITE J.M., K R.H., 1976 Overall classification as a selection criterion for improving categorically scored components of type in Holsteins J Dairy Sci., 59, 2104-2121